lib/net/net_crc_avx512.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright(c) 2020 Intel Corporation
   3  */
   4
   5
   6 #include <rte_common.h>
   7
   8 #include "net_crc.h"
   9
  10 #include <x86intrin.h>
  11
  12 /* VPCLMULQDQ CRC computation context structure */
  13 struct crc_vpclmulqdq_ctx {
  14         __m512i rk1_rk2;
  15         __m512i rk3_rk4;
  16         __m512i fold_7x128b;
  17         __m512i fold_3x128b;
  18         __m128i rk5_rk6;
  19         __m128i rk7_rk8;
  20         __m128i fold_1x128b;
  21 };
  22
  23 static struct crc_vpclmulqdq_ctx crc32_eth __rte_aligned(64);
  24 static struct crc_vpclmulqdq_ctx crc16_ccitt __rte_aligned(64);
  25
  26 static uint16_t byte_len_to_mask_table[] = {
  27         0x0000, 0x0001, 0x0003, 0x0007,
  28         0x000f, 0x001f, 0x003f, 0x007f,
  29         0x00ff, 0x01ff, 0x03ff, 0x07ff,
  30         0x0fff, 0x1fff, 0x3fff, 0x7fff,
  31         0xffff};
  32
  33 static const uint8_t shf_table[32] __rte_aligned(16) = {
  34         0x00, 0x81, 0x82, 0x83, 0x84, 0x85, 0x86, 0x87,
  35         0x88, 0x89, 0x8a, 0x8b, 0x8c, 0x8d, 0x8e, 0x8f,
  36         0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07,
  37         0x08, 0x09, 0x0a, 0x0b, 0x0c, 0x0d, 0x0e, 0x0f
  38 };
  39
  40 static const uint32_t mask[4] __rte_aligned(16) = {
  41         0xffffffff, 0xffffffff, 0x00000000, 0x00000000
  42 };
  43
  44 static const uint32_t mask2[4] __rte_aligned(16) = {
  45         0x00000000, 0xffffffff, 0xffffffff, 0xffffffff
  46 };
  47
  48 static __rte_always_inline __m512i
  49 crcr32_folding_round(__m512i data_block, __m512i precomp, __m512i fold)
  50 {
  51         __m512i tmp0, tmp1;
  52
  53         tmp0 = _mm512_clmulepi64_epi128(fold, precomp, 0x01);
  54         tmp1 = _mm512_clmulepi64_epi128(fold, precomp, 0x10);
  55
  56         return _mm512_ternarylogic_epi64(tmp0, tmp1, data_block, 0x96);
  57 }
  58
  59 static __rte_always_inline __m128i
  60 crc32_fold_128(__m512i fold0, __m512i fold1,
  61         const struct crc_vpclmulqdq_ctx *params)
  62 {
  63         __m128i res, res2;
  64         __m256i a;
  65         __m512i tmp0, tmp1, tmp2, tmp3;
  66         __m512i tmp4;
  67
  68         tmp0 = _mm512_clmulepi64_epi128(fold0, params->fold_7x128b, 0x01);
  69         tmp1 = _mm512_clmulepi64_epi128(fold0, params->fold_7x128b, 0x10);
  70
  71         res = _mm512_extracti64x2_epi64(fold1, 3);
  72         tmp4 = _mm512_maskz_broadcast_i32x4(0xF, res);
  73
  74         tmp2 = _mm512_clmulepi64_epi128(fold1, params->fold_3x128b, 0x01);
  75         tmp3 = _mm512_clmulepi64_epi128(fold1, params->fold_3x128b, 0x10);
  76
  77         tmp0 = _mm512_ternarylogic_epi64(tmp0, tmp1, tmp2, 0x96);
  78         tmp0 = _mm512_ternarylogic_epi64(tmp0, tmp3, tmp4, 0x96);
  79
  80         tmp1 = _mm512_shuffle_i64x2(tmp0, tmp0, 0x4e);
  81
  82         a = _mm256_xor_si256(*(__m256i *)&tmp1, *(__m256i *)&tmp0);
  83         res = _mm256_extracti64x2_epi64(a, 1);
  84         res2 = _mm_xor_si128(res, *(__m128i *)&a);
  85
  86         return res2;
  87 }
  88
  89 static __rte_always_inline __m128i
  90 last_two_xmm(const uint8_t *data, uint32_t data_len, uint32_t n, __m128i res,
  91         const struct crc_vpclmulqdq_ctx *params)
  92 {
  93         uint32_t offset;
  94         __m128i res2, res3, res4, pshufb_shf;
  95
  96         const uint32_t mask3[4] __rte_aligned(16) = {
  97                    0x80808080, 0x80808080, 0x80808080, 0x80808080
  98         };
  99
 100         res2 = res;
 101         offset = data_len - n;
 102         res3 = _mm_loadu_si128((const __m128i *)&data[n+offset-16]);
 103
 104         pshufb_shf = _mm_loadu_si128((const __m128i *)
 105                         (shf_table + (data_len-n)));
 106
 107         res = _mm_shuffle_epi8(res, pshufb_shf);
 108         pshufb_shf = _mm_xor_si128(pshufb_shf,
 109                         _mm_load_si128((const __m128i *) mask3));
 110         res2 = _mm_shuffle_epi8(res2, pshufb_shf);
 111
 112         res2 = _mm_blendv_epi8(res2, res3, pshufb_shf);
 113
 114         res4 = _mm_clmulepi64_si128(res, params->fold_1x128b, 0x01);
 115         res = _mm_clmulepi64_si128(res, params->fold_1x128b, 0x10);
 116         res = _mm_ternarylogic_epi64(res, res2, res4, 0x96);
 117
 118         return res;
 119 }
 120
 121 static __rte_always_inline __m128i
 122 done_128(__m128i res, const struct crc_vpclmulqdq_ctx *params)
 123 {
 124         __m128i res1;
 125
 126         res1 = res;
 127
 128         res = _mm_clmulepi64_si128(res, params->rk5_rk6, 0x0);
 129         res1 = _mm_srli_si128(res1, 8);
 130         res = _mm_xor_si128(res, res1);
 131
 132         res1 = res;
 133         res = _mm_slli_si128(res, 4);
 134         res = _mm_clmulepi64_si128(res, params->rk5_rk6, 0x10);
 135         res = _mm_xor_si128(res, res1);
 136
 137         return res;
 138 }
 139
 140 static __rte_always_inline uint32_t
 141 barrett_reduction(__m128i data64, const struct crc_vpclmulqdq_ctx *params)
 142 {
 143         __m128i tmp0, tmp1;
 144
 145         data64 =  _mm_and_si128(data64, *(const __m128i *)mask2);
 146         tmp0 = data64;
 147         tmp1 = data64;
 148
 149         data64 = _mm_clmulepi64_si128(tmp0, params->rk7_rk8, 0x0);
 150         data64 = _mm_ternarylogic_epi64(data64, tmp1, *(const __m128i *)mask,
 151                         0x28);
 152
 153         tmp1 = data64;
 154         data64 = _mm_clmulepi64_si128(data64, params->rk7_rk8, 0x10);
 155         data64 = _mm_ternarylogic_epi64(data64, tmp1, tmp0, 0x96);
 156
 157         return _mm_extract_epi32(data64, 2);
 158 }
 159
 160 static __rte_always_inline void
 161 reduction_loop(__m128i *fold, int *len, const uint8_t *data, uint32_t *n,
 162         const struct crc_vpclmulqdq_ctx *params)
 163 {
 164         __m128i tmp, tmp1;
 165
 166         tmp = _mm_clmulepi64_si128(*fold, params->fold_1x128b, 0x1);
 167         *fold = _mm_clmulepi64_si128(*fold, params->fold_1x128b, 0x10);
 168         *fold = _mm_xor_si128(*fold, tmp);
 169         tmp1 = _mm_loadu_si128((const __m128i *)&data[*n]);
 170         *fold = _mm_xor_si128(*fold, tmp1);
 171         *n += 16;
 172         *len -= 16;
 173 }
 174
 175 static __rte_always_inline uint32_t
 176 crc32_eth_calc_vpclmulqdq(const uint8_t *data, uint32_t data_len, uint32_t crc,
 177         const struct crc_vpclmulqdq_ctx *params)
 178 {
 179         __m128i res, d, b;
 180         __m512i temp, k;
 181         __m512i qw0 = _mm512_set1_epi64(0), qw1, qw2, qw3;
 182         __m512i fold0, fold1, fold2, fold3;
 183         __mmask16 mask;
 184         uint32_t n = 0;
 185         int reduction = 0;
 186
 187         /* Get CRC init value */
 188         b = _mm_cvtsi32_si128(crc);
 189         temp = _mm512_castsi128_si512(b);
 190
 191         if (data_len > 255) {
 192                 fold0 = _mm512_loadu_si512((const __m512i *)data);
 193                 fold1 = _mm512_loadu_si512((const __m512i *)(data+64));
 194                 fold2 = _mm512_loadu_si512((const __m512i *)(data+128));
 195                 fold3 = _mm512_loadu_si512((const __m512i *)(data+192));
 196                 fold0 = _mm512_xor_si512(fold0, temp);
 197
 198                 /* Main folding loop */
 199                 k = params->rk1_rk2;
 200                 for (n = 256; (n + 256) <= data_len; n += 256) {
 201                         qw0 = _mm512_loadu_si512((const __m512i *)&data[n]);
 202                         qw1 = _mm512_loadu_si512((const __m512i *)
 203                                         &(data[n+64]));
 204                         qw2 = _mm512_loadu_si512((const __m512i *)
 205                                         &(data[n+128]));
 206                         qw3 = _mm512_loadu_si512((const __m512i *)
 207                                         &(data[n+192]));
 208                         fold0 = crcr32_folding_round(qw0, k, fold0);
 209                         fold1 = crcr32_folding_round(qw1, k, fold1);
 210                         fold2 = crcr32_folding_round(qw2, k, fold2);
 211                         fold3 = crcr32_folding_round(qw3, k, fold3);
 212                 }
 213
 214                 /* 256 to 128 fold */
 215                 k = params->rk3_rk4;
 216                 fold0 = crcr32_folding_round(fold2, k, fold0);
 217                 fold1 = crcr32_folding_round(fold3, k, fold1);
 218
 219                 res = crc32_fold_128(fold0, fold1, params);
 220
 221                 reduction = 240 - ((n+256)-data_len);
 222
 223                 while (reduction > 0)
 224                         reduction_loop(&res, &reduction, data, &n,
 225                                         params);
 226
 227                 reduction += 16;
 228
 229                 if (n != data_len)
 230                         res = last_two_xmm(data, data_len, n, res,
 231                                         params);
 232         } else {
 233                 if (data_len > 31) {
 234                         res = _mm_cvtsi32_si128(crc);
 235                         d = _mm_loadu_si128((const __m128i *)data);
 236                         res = _mm_xor_si128(res, d);
 237                         n += 16;
 238
 239                         reduction = 240 - ((n+256)-data_len);
 240
 241                         while (reduction > 0)
 242                                 reduction_loop(&res, &reduction, data, &n,
 243                                                 params);
 244
 245                         if (n != data_len)
 246                                 res = last_two_xmm(data, data_len, n, res,
 247                                                 params);
 248                 } else if (data_len > 16) {
 249                         res = _mm_cvtsi32_si128(crc);
 250                         d = _mm_loadu_si128((const __m128i *)data);
 251                         res = _mm_xor_si128(res, d);
 252                         n += 16;
 253
 254                         if (n != data_len)
 255                                 res = last_two_xmm(data, data_len, n, res,
 256                                                 params);
 257                 } else if (data_len == 16) {
 258                         res = _mm_cvtsi32_si128(crc);
 259                         d = _mm_loadu_si128((const __m128i *)data);
 260                         res = _mm_xor_si128(res, d);
 261                 } else {
 262                         res = _mm_cvtsi32_si128(crc);
 263                         mask = byte_len_to_mask_table[data_len];
 264                         d = _mm_maskz_loadu_epi8(mask, data);
 265                         res = _mm_xor_si128(res, d);
 266
 267                         if (data_len > 3) {
 268                                 d = _mm_loadu_si128((const __m128i *)
 269                                                 &shf_table[data_len]);
 270                                 res = _mm_shuffle_epi8(res, d);
 271                         } else if (data_len > 2) {
 272                                 res = _mm_slli_si128(res, 5);
 273                                 goto do_barrett_reduction;
 274                         } else if (data_len > 1) {
 275                                 res = _mm_slli_si128(res, 6);
 276                                 goto do_barrett_reduction;
 277                         } else if (data_len > 0) {
 278                                 res = _mm_slli_si128(res, 7);
 279                                 goto do_barrett_reduction;
 280                         } else {
 281                                 /* zero length case */
 282                                 return crc;
 283                         }
 284                 }
 285         }
 286
 287         res = done_128(res, params);
 288
 289 do_barrett_reduction:
 290         n = barrett_reduction(res, params);
 291
 292         return n;
 293 }
 294
 295 static void
 296 crc32_load_init_constants(void)
 297 {
 298         __m128i a;
 299         /* fold constants */
 300         uint64_t c0 = 0x00000000e95c1271;
 301         uint64_t c1 = 0x00000000ce3371cb;
 302         uint64_t c2 = 0x00000000910eeec1;
 303         uint64_t c3 = 0x0000000033fff533;
 304         uint64_t c4 = 0x000000000cbec0ed;
 305         uint64_t c5 = 0x0000000031f8303f;
 306         uint64_t c6 = 0x0000000057c54819;
 307         uint64_t c7 = 0x00000000df068dc2;
 308         uint64_t c8 = 0x00000000ae0b5394;
 309         uint64_t c9 = 0x000000001c279815;
 310         uint64_t c10 = 0x000000001d9513d7;
 311         uint64_t c11 = 0x000000008f352d95;
 312         uint64_t c12 = 0x00000000af449247;
 313         uint64_t c13 = 0x000000003db1ecdc;
 314         uint64_t c14 = 0x0000000081256527;
 315         uint64_t c15 = 0x00000000f1da05aa;
 316         uint64_t c16 = 0x00000000ccaa009e;
 317         uint64_t c17 = 0x00000000ae689191;
 318         uint64_t c18 = 0x00000000ccaa009e;
 319         uint64_t c19 = 0x00000000b8bc6765;
 320         uint64_t c20 = 0x00000001f7011640;
 321         uint64_t c21 = 0x00000001db710640;
 322
 323         a = _mm_set_epi64x(c1, c0);
 324         crc32_eth.rk1_rk2 = _mm512_broadcast_i32x4(a);
 325
 326         a = _mm_set_epi64x(c3, c2);
 327         crc32_eth.rk3_rk4 = _mm512_broadcast_i32x4(a);
 328
 329         crc32_eth.fold_7x128b = _mm512_setr_epi64(c4, c5, c6, c7, c8,
 330                         c9, c10, c11);
 331         crc32_eth.fold_3x128b = _mm512_setr_epi64(c12, c13, c14, c15,
 332                         c16, c17, 0, 0);
 333         crc32_eth.fold_1x128b = _mm_setr_epi64(_mm_cvtsi64_m64(c16),
 334                         _mm_cvtsi64_m64(c17));
 335
 336         crc32_eth.rk5_rk6 = _mm_setr_epi64(_mm_cvtsi64_m64(c18),
 337                         _mm_cvtsi64_m64(c19));
 338         crc32_eth.rk7_rk8 = _mm_setr_epi64(_mm_cvtsi64_m64(c20),
 339                         _mm_cvtsi64_m64(c21));
 340 }
 341
 342 static void
 343 crc16_load_init_constants(void)
 344 {
 345         __m128i a;
 346         /* fold constants */
 347         uint64_t c0 = 0x0000000000009a19;
 348         uint64_t c1 = 0x0000000000002df8;
 349         uint64_t c2 = 0x00000000000068af;
 350         uint64_t c3 = 0x000000000000b6c9;
 351         uint64_t c4 = 0x000000000000c64f;
 352         uint64_t c5 = 0x000000000000cd95;
 353         uint64_t c6 = 0x000000000000d341;
 354         uint64_t c7 = 0x000000000000b8f2;
 355         uint64_t c8 = 0x0000000000000842;
 356         uint64_t c9 = 0x000000000000b072;
 357         uint64_t c10 = 0x00000000000047e3;
 358         uint64_t c11 = 0x000000000000922d;
 359         uint64_t c12 = 0x0000000000000e3a;
 360         uint64_t c13 = 0x0000000000004d7a;
 361         uint64_t c14 = 0x0000000000005b44;
 362         uint64_t c15 = 0x0000000000007762;
 363         uint64_t c16 = 0x00000000000081bf;
 364         uint64_t c17 = 0x0000000000008e10;
 365         uint64_t c18 = 0x00000000000081bf;
 366         uint64_t c19 = 0x0000000000001cbb;
 367         uint64_t c20 = 0x000000011c581910;
 368         uint64_t c21 = 0x0000000000010810;
 369
 370         a = _mm_set_epi64x(c1, c0);
 371         crc16_ccitt.rk1_rk2 = _mm512_broadcast_i32x4(a);
 372
 373         a = _mm_set_epi64x(c3, c2);
 374         crc16_ccitt.rk3_rk4 = _mm512_broadcast_i32x4(a);
 375
 376         crc16_ccitt.fold_7x128b = _mm512_setr_epi64(c4, c5, c6, c7, c8,
 377                         c9, c10, c11);
 378         crc16_ccitt.fold_3x128b = _mm512_setr_epi64(c12, c13, c14, c15,
 379                         c16, c17, 0, 0);
 380         crc16_ccitt.fold_1x128b = _mm_setr_epi64(_mm_cvtsi64_m64(c16),
 381                         _mm_cvtsi64_m64(c17));
 382
 383         crc16_ccitt.rk5_rk6 = _mm_setr_epi64(_mm_cvtsi64_m64(c18),
 384                         _mm_cvtsi64_m64(c19));
 385         crc16_ccitt.rk7_rk8 = _mm_setr_epi64(_mm_cvtsi64_m64(c20),
 386                         _mm_cvtsi64_m64(c21));
 387 }
 388
 389 void
 390 rte_net_crc_avx512_init(void)
 391 {
 392         crc32_load_init_constants();
 393         crc16_load_init_constants();
 394
 395         /*
 396          * Reset the register as following calculation may
 397          * use other data types such as float, double, etc.
 398          */
 399         _mm_empty();
 400 }
 401
 402 uint32_t
 403 rte_crc16_ccitt_avx512_handler(const uint8_t *data, uint32_t data_len)
 404 {
 405         /* return 16-bit CRC value */
 406         return (uint16_t)~crc32_eth_calc_vpclmulqdq(data,
 407                 data_len,
 408                 0xffff,
 409                 &crc16_ccitt);
 410 }
 411
 412 uint32_t
 413 rte_crc32_eth_avx512_handler(const uint8_t *data, uint32_t data_len)
 414 {
 415         /* return 32-bit CRC value */
 416         return ~crc32_eth_calc_vpclmulqdq(data,
 417                 data_len,
 418                 0xffffffffUL,
 419                 &crc32_eth);
 420 }